The Bayes factor is a key quantity in Bayesian statistics that compares the relative support for two competing hypotheses or models given observed data, defined as the ratio of the marginal likelihood of the data under one model to that under the other.^[1] It serves as an updating factor on prior odds, where a Bayes factor $ B_{10} = 6 $, for instance, indicates that the data are six times more likely under the alternative hypothesis $ H_1 $ than under the null hypothesis $ H_0 $.^[1] Originating in the work of Harold Jeffreys in 1935, the concept built on earlier contributions from Dorothy Wrinch and J.B.S. Haldane in the 1920s and 1930s, with Jeffreys formalizing it as a tool for scientific inference in his influential book Theory of Probability.^[1] The term "Bayes factor" itself was coined later by Robert E. Kass and Adrian E. Raftery in their 1995 paper, which popularized its use for model selection and hypothesis testing. Unlike frequentist p-values, which only assess evidence against a null hypothesis, the Bayes factor provides a symmetric measure that can quantify evidence in favor of either hypothesis, distinguishing between absence of evidence and evidence of absence.^[1] It is particularly advantageous for comparing non-nested models and is robust to optional stopping in data collection, making it suitable for sequential experimental designs.^[1] Jeffreys proposed a heuristic scale to interpret Bayes factor magnitudes as strength of evidence: values between 1 and 3 indicate "anecdotal" support for $ H_1 $, 3 to 10 offer "moderate" evidence, 10 to 30 provide "strong" evidence, 30 to 100 yield "very strong" evidence, and greater than 100 represent "extreme" evidence, with reciprocals applying for support of $ H_0 $.^[1] This scale, while subjective, has been widely adopted and refined in fields like psychology, neuroscience, and economics for model comparison tasks. Bayes factors often require numerical approximations due to the intractability of marginal likelihoods in complex models, but they remain central to Bayesian decision-making and evidence accumulation.

Mathematical Foundations

Definition

The Bayes factor is a statistical measure used in Bayesian inference to quantify the relative evidence provided by observed data for one model over another competing model. It was introduced by Harold Jeffreys as a tool for objective hypothesis testing within a Bayesian framework.^[2] Mathematically, the Bayes factor in favor of model $M_1$ over model $M_2$ given data $D$, denoted $BF_{10}$, is defined as the ratio of the marginal likelihoods under each model:

$$BF_{10} = \frac{p(D \mid M_1)}{p(D \mid M_2)}$$

The marginal likelihood $p(D \mid M)$ for a model $M$ with parameters $\theta$ is obtained by integrating the likelihood over the prior distribution of the parameters:

$$p(D \mid M) = \int p(D \mid \theta, M) \, p(\theta \mid M) \, d\theta$$

This integration averages the model's predictive performance across all plausible parameter values weighted by the prior, providing a summary of the model's overall fit to the data independent of specific parameter estimates.^[2] A common notation convention is $BF_{01} = 1 / BF_{10}$, which reverses the comparison to favor $M_0$ (often the null model) over $M_1$. The Bayes factor plays a central role in Bayesian model comparison by directly comparing the predictive adequacy of competing models based on the observed data, facilitating decisions about model selection without relying on point estimates or frequentist criteria.^[2]

Relationship to Bayes' Theorem

The Bayes factor emerges directly from Bayes' theorem as a key component in updating the probabilities of competing models based on observed data. Bayes' theorem states that the posterior probability of a model $M_i$ given data $D$ is $P(M_i | D) \propto P(D | M_i) P(M_i)$, where $P(D | M_i)$ is the marginal likelihood under the model and $P(M_i)$ is the prior probability. For two models $M_1$ and $M_2$, the ratio of posterior model probabilities, known as the posterior odds, is therefore

$$\frac{P(M_1 | D)}{P(M_2 | D)} = \frac{P(M_1)}{P(M_2)} \times \frac{P(D | M_1)}{P(D | M_2)},$$

with the second factor on the right-hand side defining the Bayes factor $BF_{12}$.^[2] This formulation demonstrates that the Bayes factor serves as a multiplier that adjusts the prior odds to yield the posterior odds, encapsulating how the data shifts belief between models.^[2] By isolating $\frac{P(D | M_1)}{P(D | M_2)}$, the Bayes factor measures the relative support for each model provided solely by the data, disentangling this evidential contribution from subjective prior beliefs about the models' plausibility.^[2] This separation allows the Bayes factor to function as an objective summary of the data's evidential value within the Bayesian updating process, applicable across diverse modeling contexts.^[2] The derivation of the Bayes factor highlights a fundamental distinction in handling point-null hypotheses versus composite models. Under a point-null hypothesis, such as $M_0: \theta = \theta_0$, the marginal likelihood $P(D | M_0)$ simplifies to the likelihood evaluated directly at the fixed parameter value, as there is no parameter uncertainty to integrate over.^[2] In contrast, for a composite model $M_1$ with parameters varying over a continuous space, $P(D | M_1)$ requires integrating the likelihood over a prior distribution on the parameters to average out uncertainty, as previously outlined in the definition of marginal likelihood.^[2] This difference affects the computational form of the Bayes factor but preserves its role in the posterior odds equation.^[2] Harold Jeffreys pioneered the application of the Bayes factor within this framework of Bayes' theorem in his 1939 monograph Theory of Probability (first edition).^[3][2]

Interpretation

Evidence Scales

The interpretation of the Bayes factor (BF) relies on standardized scales that categorize its magnitude into qualitative levels of evidence for one model (say, the alternative $M_1$) over another (say, the null $M_0$). These scales provide a heuristic framework for assessing evidential strength, though they are not universally fixed.^[4] A seminal scale was proposed by Harold Jeffreys, which divides BF values into grades based on orders of magnitude, emphasizing decisive evidence for large values. Jeffreys' scale, as commonly referenced, is as follows: This classification interprets BF$_{10} > 1$ as favoring $M_1$ over $M_0$, with the strength increasing as the value grows; reciprocally, BF$_{10} < 1$ (or BF$_{01} > 1$) supports $M_0$.^[4][5] Kass and Raftery later modified this scale to align more closely with logarithmic transformations of odds, adjusting thresholds for practicality in empirical applications and incorporating a deviance-like measure (2 ln(BF)). Their revised scale is: This adjustment extends the "very strong" category to higher values while broadening the "positive" range, facilitating interpretation in contexts like model selection.^[4] Both scales maintain the directional guideline that BF$_{10} > 1$ indicates data more compatible with $M_1$ than $M_0$, with the ratio quantifying the relative evidential support.^[4] Despite their utility, these thresholds are inherently arbitrary, serving as rough guides rather than strict cutoffs, and have varied across implementations (e.g., some adaptations use "extreme" instead of "decisive" for BF > 100).^[4] Moreover, Bayes factors are sensitive to model specification, as changes in how competing models are parameterized or nested can substantially alter the marginal likelihoods and thus the BF value, underscoring the need for careful model formulation.

Posterior Odds Connection

The Bayes factor connects directly to posterior odds through Bayes' theorem applied to model comparison. Specifically, the posterior odds in favor of model $M_1$ over model $M_2$ given data $D$ are obtained by multiplying the prior odds by the Bayes factor:

$$\frac{P(M_1 \mid D)}{P(M_2 \mid D)} = BF_{10} \times \frac{P(M_1)}{P(M_2)},$$

where $BF_{10}$ is the Bayes factor comparing $M_1$ to $M_2$. This relationship highlights the Bayes factor's role as the multiplicative update factor representing the evidence contributed solely by the data, independent of prior beliefs. Expanding this to posterior probabilities, let $\pi_1 = P(M_1)$ and $\pi_2 = P(M_2) = 1 - \pi_1$; then

$$P(M_1 \mid D) = \frac{BF_{10} \pi_1}{BF_{10} \pi_1 + \pi_2}.$$

In model selection, when prior probabilities are fixed, the Bayes factor serves as a sufficient statistic for the evidential content of the data, allowing direct quantification of how the observed data shifts belief between competing models without needing to recompute full posteriors for each prior adjustment. This makes it particularly valuable for objective comparisons, as it isolates the data's influence while priors handle subjective elements. A common default assumption in Bayes factor applications is equal prior probabilities ($\pi_1 = \pi_2 = 0.5$), which simplifies the posterior odds to equal the Bayes factor itself and the posterior probability to $P(M_1 \mid D) = \frac{BF_{10}}{1 + BF_{10}}$. This assumption rests on the premise that the models are a priori equally plausible, often justified in exploratory analyses or when domain knowledge lacks strong preferences, though it can be sensitive to model complexity if not carefully considered. In contrast to non-Bayesian approaches, where likelihood ratios compare point estimates of parameters under each model, the Bayes factor employs marginal likelihoods that integrate over parameter priors, providing a fuller evidential measure that accounts for model uncertainty.

Computation Methods

Exact Calculation

Exact calculation of the Bayes factor is possible in cases where the marginal likelihoods under each model can be derived analytically or evaluated via direct numerical methods, particularly for models with low-dimensional parameter spaces or conjugate prior distributions. These approaches avoid the need for simulation-based approximations and provide precise values, though they are limited to relatively simple model structures. In models employing conjugate priors, such as the normal distribution with known variance or the binomial distribution with a beta prior, the marginal likelihoods admit closed-form expressions, enabling straightforward computation of the Bayes factor. For instance, consider testing a point null hypothesis $ H_0: \mu = \mu_0 $ against an alternative $ H_1: \mu \sim \mathcal{N}(\mu_0, \sigma_0^2) $ for data $ x_1, \dots, x_n \iid \mathcal{N}(\mu, \sigma^2) $ with known $ \sigma^2 $. The Bayes factor $ BF_{01} $ favoring the null is given by

$$BF_{01} = \sqrt{ \frac{ \sigma_0^2 + \frac{\sigma^2}{n} }{ \frac{\sigma^2}{n} } } \exp\left( -\frac{ n (\bar{x} - \mu_0)^2 \sigma_0^2 }{ 2 \sigma^2 \left( \sigma_0^2 + \frac{\sigma^2}{n} \right) } \right).$$

This formula arises from the ratio of the normal marginal likelihood under the null to the integrated likelihood under the alternative prior.^[6] Similarly, for a binomial model testing $ H_0: p = p_0 $ against $ H_1: p \sim \text{Beta}(\alpha, \beta) $, the marginal likelihood under $ H_1 $ is the beta-binomial probability mass function, $ p(k | n, \alpha, \beta) = \binom{n}{k} \frac{ B(\alpha + k, \beta + n - k) }{ B(\alpha, \beta) } $, where $ k $ is the number of successes and $ B $ is the beta function, yielding an exact Bayes factor as the ratio to the null binomial probability.^[6] When analytical solutions are unavailable but the parameter dimensionality remains low (e.g., one or two parameters), numerical integration techniques such as Gaussian quadrature can evaluate the required integrals for the marginal likelihoods with high precision. These methods discretize the integral over the parameter space using carefully chosen nodes and weights to approximate the exact value, making them suitable for exact computation in feasible cases.^[6] For slightly more complex low-dimensional settings, Laplace approximations provide near-exact results by expanding the integrand around its mode, though they rely on asymptotic assumptions for accuracy.^[6] For nested models, where the null model is a special case of the alternative (e.g., imposing a point restriction $ \theta = \theta_0 $), the Savage-Dickey density ratio offers an exact computational shortcut under specific prior conditions. The Bayes factor $ BF_{01} $ is then

$$BF_{01} = \frac{p(\theta_0 \mid D, M_1)}{p(\theta_0 \mid M_1)},$$

provided the prior distribution for the nuisance parameters under $ M_1 $ matches that under $ M_0 $ when $ \theta \to \theta_0 $, and the posterior and prior densities are continuous at $ \theta_0 $. This ratio equates the marginal likelihoods without full integration over the alternative model.^[7] Software tools facilitate these exact methods for standard models. The R package BayesFactor implements analytical and numerical integration (via Monte Carlo with adjustable iterations for precision) to compute Bayes factors precisely for basic designs, including one-sample t-tests (equivalent to normal means with known variance under certain priors) and linear models.^[8]

Approximations and Algorithms

Computing Bayes factors exactly becomes infeasible for complex, high-dimensional models where the marginal likelihood integral cannot be evaluated analytically. Monte Carlo methods provide scalable approximations by estimating the marginal likelihood through simulation. Importance sampling draws samples from a proposal distribution to approximate the posterior, reweighting them to estimate the evidence; for instance, schemes tailored to mixture models use maximum likelihood estimates or Rao-Blackwellized dual sampling to mitigate bias from posterior mode exploration issues, enabling reliable Bayes factor computation in such settings.^[9] The harmonic mean estimator, derived from posterior samples, inverts the identity $\hat{p}(y) = \left( \frac{1}{S} \sum_{s=1}^S \frac{1}{p(y \mid \theta^{(s)})} \right)^{-1}$, where $\theta^{(s)}$ are MCMC draws from the posterior, offering a simple yet variance-prone approach to marginal likelihoods for Bayes factors.^[10] Markov chain Monte Carlo (MCMC) techniques extend these approximations for more robust estimation in nested or non-nested models. Bridge sampling leverages samples from prior and posterior distributions to estimate the normalizing constant ratio via $\hat{p}(y) = \frac{1}{S} \sum_{s=1}^S \frac{p(y \mid \theta_1^{(s)})}{q(\theta_1^{(s)} \mid y)} \cdot \frac{\int q(\theta_2 \mid y) d\theta_2}{\int \frac{p(y \mid \theta_2^{(t)})}{q(\theta_2^{(t)} \mid y)} p(\theta_2^{(t)}) d\theta_2}$, where $q$ bridges the distributions, yielding accurate Bayes factors with reduced variance compared to importance sampling alone.^[11] Thermodynamic integration approximates the marginal likelihood by integrating the expected log-likelihood along a power posterior path $\beta \in [0,1]$, $\log p(y) = \int_0^1 \mathbb{E}_{\pi(\theta \mid y^\beta)} [\log p(y \mid \theta)] d\beta$, often implemented with MCMC at discrete $\beta$ levels; this method excels for comparing phylogenetic or cognitive models, providing stable Bayes factors even in high dimensions. Recent enhancements, such as differential evolution MCMC for thermodynamic integration, further improve efficiency by requiring fewer samples per path rung, achieving convergence 5-8 times faster than standard implementations.^[12] Nested sampling is another class of algorithms for approximating marginal likelihoods, particularly effective in high-dimensional spaces. It transforms the evidence integral into a one-dimensional integral over prior mass, using sequential sampling to estimate it efficiently without tuning, as implemented in tools like MultiNest or diffuse nested sampling. This method is popular in fields like cosmology and provides reliable Bayes factors for complex models.^[13] Information criteria offer asymptotic approximations to Bayes factors without simulation. The Bayesian information criterion (BIC) estimates $\log p(y \mid M) \approx -\frac{1}{2} \mathrm{BIC} = L - \frac{k}{2} \log n$, where $L$ is the maximized log-likelihood, $k$ the number of parameters, and $n$ the sample size, deriving from Laplace's method under a unit information prior; thus, $\log \mathrm{BF}_{12} \approx \frac{\mathrm{BIC}_2 - \mathrm{BIC}_1}{2}$.^[14] This approximation holds asymptotically for large $n$ and fixed $k$, assuming regularity conditions like identifiability and correct model specification, but falters in small samples or high dimensions where the unit prior mismatches the true scenario, potentially biasing model selection.^[6] For large datasets, variational inference and integrated nested Laplace approximations (INLA) enable faster marginal likelihood estimates. Variational methods optimize a lower bound on the evidence, $\log p(y) \geq \mathbb{E}_{q(\theta)} [\log p(y \mid \theta)] - \mathrm{KL}(q(\theta) \| p(\theta))$, approximating the posterior with a tractable $q$ to derive Bayes factors in factor analysis or mixture settings, though they may underestimate evidence due to the bound's conservatism.^[15] INLA targets latent Gaussian models, combining Laplace approximations for conditional modes with numerical integration for hyperparameters to compute marginal posteriors and likelihoods efficiently; it supports Bayes factor estimation via model averaging in spatial or time-series contexts, scaling to thousands of observations without MCMC.^[16] Post-2020 advances refine these for broader applicability. Path sampling, an extension of thermodynamic integration, estimates evidence ratios by simulating paths between models, improving accuracy in non-nested comparisons for hydrological or evolutionary models.^[17] Generalized harmonic mean estimators, such as the learnt variant, employ machine learning to optimize the importance proposal from posterior samples, reducing variance by orders of magnitude and enabling scalable Bayes factors in dimensions up to $10^3$, outperforming traditional methods in speed and precision for cosmological and statistical applications.^[18]

Examples and Applications

Basic Coin Flip Example

Consider a simple hypothesis testing scenario involving coin flips to illustrate the Bayes factor. Suppose we observe data $D$ consisting of 8 heads in 10 independent flips. We compare two models: $M_0$, the null hypothesis that the coin is fair with fixed bias $\theta = 0.5$; and $M_1$, the alternative hypothesis that the coin is biased with $\theta$ following a uniform prior distribution Beta(1,1) on [0,1]. The marginal likelihood under $M_0$ is the binomial probability of the data given $\theta = 0.5$:

$$p(D \mid M_0) = \binom{10}{8} (0.5)^{10} = \frac{45}{1024} \approx 0.0439.$$

Under $M_1$, the marginal likelihood integrates the binomial likelihood over the prior:

$$p(D \mid M_1) = \int_0^1 \binom{10}{8} \theta^8 (1 - \theta)^2 \, d\theta = \binom{10}{8} \frac{B(9,3)}{B(1,1)} = \frac{45}{495} = \frac{1}{11} \approx 0.0909,$$

where $B(a,b)$ is the beta function. The Bayes factor in favor of $M_1$ over $M_0$ is then

$$BF_{10} = \frac{p(D \mid M_1)}{p(D \mid M_0)} \approx \frac{0.0909}{0.0439} \approx 2.07.$$

This calculation shows how the Bayes factor quantifies the relative evidential support for the biased coin model. According to Jeffreys' scale for interpreting Bayes factors, a value of $BF_{10}$ between 1 and 3 provides "barely worth mentioning" support for the alternative hypothesis—in this case, evidence for a biased coin. To visualize the models, consider a plot of the prior and posterior distributions under $M_1$ alongside the point mass at $\theta = 0.5$ under $M_0$. The prior under $M_1$ is flat (uniform on [0,1]). The posterior under $M_1$ is Beta(9,3), which peaks around $\theta \approx 0.75$ and shifts mass toward higher values of $\theta$ after observing 8 heads. The point mass under $M_0$ remains fixed at 0.5, highlighting the concentrated evidence for fairness versus the spread under the alternative. The relative heights of the predictive distributions at the observed data further illustrate why $M_1$ receives more support here.

Model Comparison in Regression

In linear regression analysis, Bayes factors enable the comparison of nested or non-nested models by quantifying the relative evidence provided by the data for each model, facilitating decisions on predictor inclusion. A typical scenario involves contrasting a null model, which includes only an intercept, against an alternative model incorporating a single predictor. For instance, consider simulated data with 50 observations where the response variable follows a linear relationship with the predictor under a moderate effect size, such as a standardized regression coefficient of approximately 0.5, reflecting realistic conditions in empirical research. This setup allows researchers to evaluate whether the predictor explains a meaningful portion of the variance beyond chance. To compute the Bayes factor in this regression context, Zellner's g-prior is commonly employed for the regression coefficients, specifying a multivariate normal distribution centered at zero with covariance proportional to the inverse of the design matrix scaled by the hyperparameter g, which tunes the prior's informativeness. The marginal likelihood under this prior admits a closed-form expression, enabling direct calculation of the Bayes factor between models. For practical implementation, approximations such as the Bayesian Information Criterion (BIC) can be used, where the difference in BIC scores between the alternative and null models approximates twice the log Bayes factor, offering computational efficiency for initial assessments. Alternatively, Markov Chain Monte Carlo (MCMC) methods provide more precise estimates by sampling from the posterior, as facilitated by tools like the BayesFactor R package, which defaults to a g-prior with g integrated over a hyperprior for robustness. In the simulated example, this yields a hypothetical BF_{10} = 5, signifying substantial evidence favoring the inclusion of the predictor according to established interpretive guidelines.^[19][2][20] Bayes factors find practical application in fields like psychology, where they support model comparisons in replication studies to assess the reliability of predictor effects across datasets, and in genetics, aiding the evaluation of whether genetic markers enhance linear models of quantitative traits in association analyses. For illustration, applying this framework to the Iris dataset—predicting sepal length from petal length—demonstrates how Bayes factors can quantify evidence for predictor utility in a real, multivariate biological context. However, results in regression settings are sensitive to prior specifications; for example, small values of g impose tighter shrinkage on coefficients, potentially reducing evidence for the alternative model, while larger g values approach non-informative priors and may inflate Bayes factors, underscoring the need for careful prior justification based on domain knowledge.^[21][19]

Historical Development

Origins and Key Contributors

The concept of the Bayes factor emerged in the early 20th century as a tool for Bayesian hypothesis testing, building on foundational principles of probability updating. Building on earlier contributions from Dorothy Wrinch and J.B.S. Haldane in the 1920s and 1930s, which introduced ideas of evidence accumulation and likelihood-based hypothesis testing.^[22] Harold Jeffreys first introduced the Bayes factor in his 1935 paper "Some tests of significance, treated by the theory of probability," which he further developed in his 1939 book Theory of Probability, where he defined it as the ratio of the likelihood of data under competing hypotheses to quantify evidence in favor of one model over another.^[23] Jeffreys applied this approach to geophysical problems, such as testing hypotheses about the rigidity of the Earth's core against seismic data, demonstrating its utility in scientific inference beyond subjective priors.^[23] Although Jeffreys's work was explicitly Bayesian, it drew indirect influence from Sir Ronald Fisher's development of likelihood ratios in the 1920s, which provided a non-Bayesian framework for comparing models based on data likelihoods alone.^[24] Fisher's likelihood ratio tests emphasized evidential strength without prior probabilities, laying groundwork that Bayesians later extended by incorporating priors to form the full Bayes factor.^[24] In the 1950s, I. J. Good extended these ideas by formalizing "odds factors" as measures of evidential weight, particularly in contexts linking probability to information theory, where he explored how data updates prior odds through logarithmic transformations akin to entropy.^[25] Good's contributions emphasized the Bayes factor's role in weighing evidence objectively, influencing its application in decision theory and cryptography-related statistical problems.^[25] Key advancements in the 1960s and 1970s highlighted the Bayes factor's advantages over frequentist methods, notably in the 1963 paper by Ward Edwards, Harold Lindman, and Leonard J. Savage, which advocated Bayesian inference for psychological research by contrasting posterior odds derived from Bayes factors with p-values.^[26] This work argued that Bayes factors provide a direct measure of evidential support for hypotheses, addressing limitations in significance testing by integrating prior beliefs with data.^[26] Early recognition of computational challenges in Bayes factor evaluation came from Dennis Lindley in his 1957 paper "A Statistical Paradox," which illustrated difficulties in calculating posterior probabilities for point null hypotheses under vague priors, especially in large-sample settings where integrals become intractable.^[27] Lindley's analysis underscored the paradox where significant frequentist evidence fails to shift Bayesian posteriors substantially, pointing to the need for careful prior specification and numerical methods to make Bayes factors practical.^[27]

Evolution in Statistical Practice

The adoption of Bayes factors in statistical practice gained significant momentum in the 1980s and 1990s, building on Harold Jeffreys' foundational theoretical work from the mid-20th century.^[2] A pivotal advancement came with the 1995 paper by Robert E. Kass and Adrian E. Raftery, which provided a comprehensive review and standardization of Bayes factor interpretation and computation, particularly through practical guidelines for assessing evidence strength in model comparison. Published in the Journal of the American Statistical Association, this work addressed computational challenges and proposed interpretive scales (e.g., Bayes factors between 1 and 3 indicating "barely worth mentioning" evidence), making the method more accessible for applied researchers across fields like genetics and ecology.^[2] The 2000s marked a rise in practical implementation through accessible software tools, facilitating broader use in empirical research. The BayesFactor R package, developed by Richard D. Morey and Jeffrey N. Rouder and first released in 2012, enabled straightforward computation of Bayes factors for common designs such as t-tests, ANOVA, and regression, with default priors based on JASP guidelines for reproducibility.^[28] Complementing this, JASP—an open-source graphical interface launched around 2015 by Eric-Jan Wagenmakers and collaborators—integrated Bayes factor analyses with user-friendly defaults, promoting its adoption in psychological and social science workflows by automating prior specification and output visualization. These tools democratized Bayes factor use, shifting it from theoretical discussions to routine hypothesis testing in software environments like R.^[8] In the 2010s, Bayes factors experienced a surge in psychology and social sciences, particularly following the reproducibility crisis highlighted by low replication rates in landmark studies around 2011–2015.^[29] Eric-Jan Wagenmakers played a key role in this advocacy, promoting Bayes factors as a superior alternative to p-values for quantifying evidence and addressing selective reporting biases, as detailed in his 2016 analysis of the Reproducibility Project: Psychology, where Bayes factors provided nuanced assessments of original versus replication findings. This period saw increased publications and guidelines emphasizing Bayes factors for robust inference, with journals like Psychological Methods featuring tutorials on their application to mitigate crisis-driven skepticism toward null hypothesis significance testing.^[30] From 2020 to 2025, trends have focused on integrating Bayes factors with machine learning techniques to handle big data challenges, such as scalable model selection in high-dimensional settings. Methods like deep Bayes factors, proposed in 2023, leverage neural networks to approximate marginal likelihoods efficiently for large datasets, enabling applications in genomics and predictive modeling where traditional computations falter. Concurrently, efforts to resolve Lindley's paradox—where Bayes factors overly favor null hypotheses in large samples—have advanced through objective Bayesian approaches and interval-null testing, with 2021 and 2025 works demonstrating prior adjustments that align Bayesian evidence with practical effect sizes in big data contexts.^[31] An influential approximation enhancing this evolution has been the Bayesian Information Criterion (BIC), originally formulated by Gideon Schwarz in 1978 as a frequentist tool for model dimensionality estimation but later recognized as an asymptotic approximation to the Bayes factor under certain priors. Its Bayesian reinterpretation in the 1990s and 2000s, as elaborated by Kass and Raftery, has made BIC a computationally efficient proxy for Bayes factors in large-sample scenarios, widely adopted in software for quick model comparisons without full posterior sampling.

Comparisons and Limitations

Versus Frequentist Approaches

The Bayes factor (BF) differs fundamentally from frequentist approaches like p-values in its conceptualization of evidence. Whereas p-values measure the probability of observing data as extreme or more extreme than the observed under the null hypothesis, assuming long-run frequency properties, they often result in binary decisions (e.g., reject at α = 0.05) that do not quantify support for the alternative. In contrast, the BF provides a continuous measure of relative evidence between two hypotheses, H₀ and H₁, defined as the ratio of their marginal likelihoods, explicitly incorporating prior distributions to update beliefs via posterior odds. This allows the BF to assess evidence in favor of either hypothesis, avoiding the asymmetry inherent in p-values that only tests against the null.^[2][32] A striking example of these differences is Lindley's paradox, which arises in large-sample settings where a small deviation from the null yields a statistically significant p-value (e.g., rejecting H₀), but the BF may strongly favor the null if the prior on the alternative hypothesis is sufficiently diffuse. This occurs because the BF averages evidence over the entire parameter space under each model, diluting support for the alternative as sample size grows without strong effect magnitude, whereas p-values focus on tail probabilities that become sensitive to minor discrepancies. The paradox underscores calibration issues, where frequentist procedures amplify evidence against the null in high-power scenarios, while Bayesian methods require substantive prior specification to align conclusions. Relative to likelihood ratio tests (LRT), which compare maximized likelihoods between models and can favor overly complex nested models by overfitting noise in the data, the BF integrates out nuisance parameters using priors, yielding marginal likelihoods that inherently penalize excessive complexity. This averaging prevents the LRT's tendency to select models that fit sample idiosyncrasies without generalizing, providing a more stable comparison especially when models are nested. Empirical calibrations further highlight the BF's conservatism; for instance, Sellke et al. (2001) showed that a p-value of 0.05 corresponds to a maximum Bayes factor in favor of the alternative (BF_{10}) of about 2.5:1 (or minimum BF_{01} of about 0.4:1), indicating at best modest evidence against the null, while p=0.01 corresponds to a maximum BF_{10} of approximately 7.9:1, revealing how p-values often overestimate evidence against the null compared to BFs.^[2][33] Hybrid methods bridge these paradigms by embedding BFs within frequentist-inspired frameworks via objective Bayes approaches, which select non-informative priors to achieve properties like frequentist coverage while retaining Bayesian coherence. For example, reference priors ensure BFs approximate classical confidence intervals in limiting cases, facilitating their use in settings demanding both evidential quantification and error-rate control.^[34]

Common Criticisms

One major criticism of the Bayes factor is its sensitivity to the choice of prior distributions, which can lead to varying conclusions depending on subjective or arbitrary prior specifications. For instance, the use of unit information priors, intended to provide an objective benchmark by incorporating the information content of a single observation, has been debated for potentially overstating or understating evidence in certain models. Sensitivity analyses are often recommended to assess how alterations in priors affect the Bayes factor, as even modest changes can substantially shift the results.^[35][36] Another limitation is the computational intractability of exact Bayes factor calculations, particularly in high-dimensional settings where the marginal likelihood becomes difficult to evaluate due to the curse of dimensionality. Even approximation methods, such as those based on variational inference or Monte Carlo sampling, can introduce biases in such scenarios, requiring extensive computational resources for reliable estimates. These approximations serve as partial solutions but do not fully resolve the challenges in complex models.^[37][38] Bayes factors are also prone to misinterpretation, with researchers sometimes treating them as direct probabilities of a hypothesis rather than ratios of marginal likelihoods, which can foster overconfidence in conclusions. This risk is heightened in applied research, where common errors include conflating the Bayes factor with posterior probabilities or ignoring its dependence on model assumptions.^[39][40] A related issue is base rate neglect, where the Bayes factor is reported in isolation without integrating prior probabilities, potentially misleading inferences in scenarios with low base rates for the hypotheses under consideration. This occurs because the Bayes factor represents only the update from the data (likelihood ratio), and neglecting the prior odds can exaggerate evidence against intuitive base rates.^[41] In recent critiques from the 2020s, particularly within reproducibility initiatives in psychology and statistics, the Bayes factor has faced scrutiny for over-reliance as a decision tool, prompting alternatives like the region of practical equivalence (ROPE) for equivalence testing. These critiques highlight how Bayes factors may not adequately address practical null hypotheses or parameter equivalence, leading to calls for more robust Bayesian estimation approaches over hypothesis testing via Bayes factors. Recent studies from 2024-2025 have further highlighted practical misuses, including selective application of Bayes factors that overestimates evidence for the null and biases in estimates for factorial designs, reinforcing calls for careful implementation in applied research.^{[42][43][44][45]}